kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

Improvements to splitter

Browse Source
This commit is contained in:
James Betker 2021-09-09 23:34:56 -06:00
parent 742f9b4010
commit 76e2c497f7
2 changed files with 15 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
codes/models/spleeter/separator.py
View File

@ -29,4 +29,4 @@ class Separator:
return { return {
'vocals': res[0].cpu().numpy(), 'vocals': res[0].cpu().numpy(),
'accompaniment': res[1].cpu().numpy() 'accompaniment': res[1].cpu().numpy()
} }

27
codes/scripts/audio/preparation/spleeter_split_voice_and_background_2.py
View File

@ -14,6 +14,7 @@ def main():
src_dir = 'F:\\split\\podcast-dump0' src_dir = 'F:\\split\\podcast-dump0'
output_dir = 'F:\\tmp\\out' output_dir = 'F:\\tmp\\out'
output_dir_bg = 'F:\\tmp\\bg' output_dir_bg = 'F:\\tmp\\bg'
output_dir_reject = 'F:\\tmp\\rejected'
output_sample_rate=22050 output_sample_rate=22050
batch_size=24 batch_size=24
@ -28,34 +29,34 @@ def main():
sep = separator.separate(waves) sep = separator.separate(waves)
for j in range(sep['vocals'].shape[0]): for j in range(sep['vocals'].shape[0]):
vocals = sep['vocals'][j] wave = waves[j].cpu().numpy()[:durations[j]]
bg = sep['accompaniment'][j] vocals = sep['vocals'][j][:durations[j]]
vmax = np.abs(vocals).mean() bg = sep['accompaniment'][j][:durations[j]]
bmax = np.abs(bg).mean() vmax = np.abs(vocals[output_sample_rate:-output_sample_rate]).mean()
bmax = np.abs(bg[output_sample_rate:-output_sample_rate]).mean()
# Only output to the "good" sample dir if the ratio of background noise to vocal noise is high enough. # Only output to the "good" sample dir if the ratio of background noise to vocal noise is high enough.
ratio = vmax / (bmax+.0000001) ratio = vmax / (bmax+.0000001)
if ratio >= 25: # These values were derived empirically if ratio >= 4: # These values were derived empirically
od = output_dir od = output_dir
out_sound = waves[j].cpu().numpy() out_sound = wave
elif ratio <= 1: elif ratio <= 2:
od = output_dir_bg od = output_dir_bg
out_sound = bg out_sound = bg
else: else:
continue print(f"Reject {paths[j]}: {ratio}")
od = output_dir_reject
out_sound = wave
# Strip out channels. # Strip out channels.
if len(out_sound.shape) > 1: if len(out_sound.shape) > 1:
out_sound = out_sound[:, 0] # Just use the first channel. out_sound = out_sound[:, 0] # Just use the first channel.
# Resize to true duration
out_sound = out_sound[:durations[j]]
# Compile an output path. # Compile an output path.
path = paths[j] path = paths[j]
reld = os.path.relpath(os.path.dirname(path), src_dir) reld = str(os.path.relpath(os.path.dirname(path), src_dir)).strip()
os.makedirs(os.path.join(od, reld), exist_ok=True) os.makedirs(os.path.join(od, reld), exist_ok=True)
relp = os.path.relpath(path, src_dir) output_path = os.path.join(od, reld, os.path.basename(path))
output_path = os.path.join(od, relp)
wavfile.write(output_path, output_sample_rate, out_sound) wavfile.write(output_path, output_sample_rate, out_sound)